Having preprocessed some mean-normalized data by additionally performing feature scaling on all input features to fit an interval, say [-1,1] (possibly by dividing each data entry by the range), and having trained certain parameters using this data-set to output certain "predictions". If I were to use these parameters on new cases, I would need to use the same method to scale (preprocess) the new data (divide by the same number - possibly range of the training set).
Coming to the point, although this new data would come from a similar distribution as the training set, there is a possibility that it would not scale to [-1,1]. How would we handle such situations? Is it "okay" for the new data not to scale to [-1,1]?
I'm fairly new to Machine Learning and Statistical Regression and my question may reflect my experience (or the lack thereof) with the subject, I sincerely apologize for the possibility of this being a naive query.
[link] [9 comments]